VisPhyWorld: Probing Physical Reasoning via Code-Driven Video Reconstruction

2026-02-09

AI Summary

This work addresses the limitation of existing evaluation methods for multimodal large language models, which predominantly rely on discriminative tasks and fail to effectively assess genuine physical reasoning capabilities. To this end, we propose a novel evaluation framework based on executable code generation, requiring models to produce runnable physics simulation code from visual inputs to reconstruct videos and verify their dynamical plausibility. Our approach decouples physical reasoning from rendering for the first time, enabling direct assessment of a model’s physical understanding through generated simulations that are verifiable, editable, and falsifiable. We introduce VisPhyBench, a benchmark comprising 108 physical templates, along with a systematic protocol to evaluate both appearance and dynamics consistency. Experiments reveal that while state-of-the-art models excel in semantic understanding, they exhibit significant deficiencies in inferring physical parameters and ensuring dynamic coherence; our pipeline successfully generates valid reconstruction videos in 97.7% of scenarios.

Technology Category

• Computer Vision: Low Level & Physics-based Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Evaluating whether Multimodal Large Language Models (MLLMs) genuinely reason about physical dynamics remains challenging. Most existing benchmarks rely on recognition-style protocols such as Visual Question Answering (VQA) and Violation of Expectation (VoE), which can often be answered without committing to an explicit, testable physical hypothesis. We propose VisPhyWorld, an execution-based framework that evaluates physical reasoning by requiring models to generate executable simulator code from visual observations. By producing runnable code, the inferred world representation is directly inspectable, editable, and falsifiable. This separates physical reasoning from rendering. Building on this framework, we introduce VisPhyBench, comprising 209 evaluation scenes derived from 108 physical templates and a systematic protocol that evaluates how well models reconstruct appearance and reproduce physically plausible motion. Our pipeline produces valid reconstructed videos in 97.7% on the benchmark. Experiments show that while state-of-the-art MLLMs achieve strong semantic scene understanding, they struggle to accurately infer physical parameters and to simulate consistent physical dynamics.

Problem

• physical reasoning
• Multimodal Large Language Models
• benchmark evaluation
• executable simulation
• visual understanding

Innovation

• physical reasoning
• executable code generation
• multimodal LLMs
• video reconstruction
• simulation-based evaluation